Video-RLM: Recursive Language Models for Long Videos

๐Ÿ“‘ ๊ฐœ์š” (Overview)

๐Ÿ“ฝ๏ธ ํ•ต์‹ฌ ๊ธฐ์ˆ  ์•„ํ‚คํ…์ฒ˜ (Key Architecture)

MIT CSAIL ์—ฐ๊ตฌ์ง„์ด ์ œ์•ˆํ•œ RLM(Recursive Language Models) ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋น„๋””์˜ค ๋„๋ฉ”์ธ์œผ๋กœ ํ™•์žฅํ•œ ์—ฐ๊ตฌ์ž…๋‹ˆ๋‹ค.

1. Master-Worker ์•„ํ‚คํ…์ฒ˜

  • Master ๋ชจ๋ธ: ๋น„๋””์˜ค ์ „์ฒด์˜ ๋งฅ๋ฝ์„ ํŒŒ์•…ํ•˜๊ณ  ํƒ์ƒ‰ ์ „๋žต์„ ์ˆ˜๋ฆฝํ•ฉ๋‹ˆ๋‹ค.
  • Worker ๋ชจ๋ธ: Master๋กœ๋ถ€ํ„ฐ ํ• ๋‹น๋ฐ›์€ ํŠน์ • ๋น„๋””์˜ค ๊ตฌ๊ฐ„์—์„œ ์†์‹ค ์—†๋Š” ์‹œ๊ฐ์  ์ฆ๊ฑฐ๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค.
  • Master์™€ Worker๊ฐ€ ๋ณ‘๋ ฌ๋กœ ์ž‘๋™ํ•˜๋ฉฐ ์ •๋ณด๋ฅผ ์žฌ๊ท€์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค.

2. ๋กœ๊ทธ ๋‹จ์œ„ ๊ณ„์‚ฐ ์ตœ์ ํ™”

  • ๋น„๋””์˜ค ๊ธธ์ด๊ฐ€ ๊ธธ์–ด์ง์— ๋”ฐ๋ผ ๊ณ„์‚ฐ๋Ÿ‰์ด ์„ ํ˜•์ด ์•„๋‹Œ ๋กœ๊ทธ(logarithmic) ๋‹จ์œ„๋กœ ์ฆ๊ฐ€ํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
  • ์ด๋ฅผ ํ†ตํ•ด ์ˆ˜ ์‹œ๊ฐ„ ๋ถ„๋Ÿ‰์˜ ๋งค์šฐ ๊ธด ์˜์ƒ์—์„œ๋„ ํšจ์œจ์ ์ธ ์ถ”๋ก ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ“ˆ ์ฃผ์š” ๊ธฐ์—ฌ ๋ฐ ์„ฑ๊ณผ (Contributions & Results)

  • ํšจ์œจ์„ฑ: ์ดˆ์žฅ๊ธฐ ๋ฌธ๋งฅ ์ฒ˜๋ฆฌ ์‹œ ๋ฐœ์ƒํ•˜๋Š” โ€˜์ปจํ…์ŠคํŠธ ๋ง๊ฐ(Context Rot)โ€™ ๋ฌธ์ œ๋ฅผ ์‹œ์Šคํ…œ ์•„ํ‚คํ…์ฒ˜์ ์œผ๋กœ ํ•ด๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค.
  • ์—์ด์ „ํŠธ ํ†ตํ•ฉ: recursive-decomposition-skill๊ณผ ์—ฐ๋™ํ•˜์—ฌ ๋ณต์žกํ•œ ์‹œ๊ฐ์  ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•˜๋Š” ์—์ด์ „ํŠธ์˜ ํ•ต์‹ฌ ์—”์ง„์œผ๋กœ ํ™œ์šฉ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.

Created on: 2026-03-20